AIDesign(tm)

Absolutely love the Midjourney --sref + personalize code combo culture that has been going on lately especially the combos shared by Tatiana Tsiguleva! Here is my first (successful) attempt training a LoRA on purely synthetic images from MJ --sref combo shared by Tatiana - love this style. Finetune testing: --sref 680572301 3322162017 3739894080 --sw 1000 --stylize 1000 --personalize kzilt9y #TrainingLoRA Link to Tatiana's post: https://lnkd.in/gQNSmDqA (personal studies unrelated to practice) #ai #aiarchitecture #architecture #design #midjourney #flux

From Simple to Stunning 🌊🏡 Starting with a basic 3D model, we used AI tools like Midjourney, ComfyUI, and Runway Gen2 to bring this architectural visualization to life. The journey from a simple floor plan to a hyper-realistic render was smooth and fast, showcasing the incredible power of AI in design. These tools are game-changers for architects, making it easier to visualize projects with stunning realism. Swipe through to see the transformation! #ArchitecturalVisualization #AIArt #Midjourney #ComfyUI #RunwayGen2 #3DModeling #AIInDesign #ArchitectureLovers #DesignInspiration #DigitalArt #VisualStorytelling #ArchitectureDaily #Rendering #CreativeProcess #ModernArchitecture #DesignInnovation

I began my career in graphic design when I was ten years old, creating greeting cards, business cards, and posters. Over the past thirty-plus years, I've watched the design industry evolve from manual methods to digital platforms, and now into the era of AI. My background in advertising, illustration, and graphic design has allowed me to approach technology development with a creative mindset. Whether it’s programming or prompt engineering, I apply design thinking to everything I do, proving that creativity isn’t limited to art—it can be woven into the fabric of any discipline. What excites me most today is the rise of AI-based graphic design. This new wave of technology allows us to apply the same iterative, creative approach I’ve always valued but at a scale and speed we couldn’t have imagined just a few years ago. We can now test and refine endless design variations in minutes, where previously it took hours. AI is transforming graphic design, much like desktop publishing once revolutionized the field by shifting us from letterpresses to digital design tools. But this evolution doesn’t mean traditional design skills are obsolete. Rather, it changes the role of the designer. We’re becoming creative directors, orchestrating AI tools to bring our visions to life. AI doesn’t eliminate creativity—it accelerates it. As we move into this next generation of design, human intuition and a creative touch will remain essential to guide AI in producing meaningful, impactful work.

An Object is Worth 64x64 Pixels: Generating 3D Objects via Image Diffusion. Basically, think of the "omages" approach as akin to #MaterialID. Indeed, that's what this looks like when the algorithm breaks down the 3D into discrete geometry. Most #TextTo3D suffers from making any resulting #3Dmodel as a single mesh. Depending on what that model is intended for, this may or may not be useful. For example, the project page refers to books on a shelf that are, ultimately, a clump of books as a single 3D object, not individual books. The Omage technique can take complex shapes and break them down into a single 64x64 pixel image, making the 3D shapes into a 2D format (that alone is difficult to grasp)! By converting these Object + Images into "Omages", the result is information that is BOTH a #pbr material and mesh geometry! Is it perfect? No. It's a first-gen approach to a challenge, but check out the takeaways below for more detail. This is a fascinating way of encoding complex and irregular 3D information. I'm curious as to whether #animation might effectively be encoded as a string of sequential 64x64-bit PNGs, which could, conceivably, be compressed ever further? Fascinatingly, to increase the object resolution, one need only rescale the omage. Curioser and curioser. Project: https://omages.github.io/ Paper: https://lnkd.in/gQ3w7PNG Code: https://lnkd.in/gt8UrW7g TAKEAWAYS: 1. Omage encodes geometry as a (R,R,3+1) image, which is essentially a 16-bit RGBA PNG! 2. PBR material is encoded in another (R,R,8) image. So three PNGs will give you a realistic 3D object. 3. You can use image generation models to generate 3D objects, one image for one object! 4. Discrete patch structures emerge out of continuous noise during the denoising process. 5. Patch segmentation comes naturally from 2D disconnected components, no instance labels needed! 6. Irregular connectivity & complex topology? No worries, all encoded in a regular image. 7. Change your object resolution by just rescaling the omage! 8. Generated shapes come with UV maps—no unwrapping required.

Finally managed to find some time to start to learn how to train LORA’s… essentially this is a process that allows you to fine tune a text to image model on your own images so you can more effectively direct the aesthetic qualities of your outputs, or effect the subject (person, product…) in those images, or both. It’s not a technically difficult process - however there are many variables to consider from what to include in the data set (images that you’re training with), how to caption the images, through to various parameters that can be set during the training process (steps, learning rate etc). Each variable has an impact on the way the LORA affects the final image output. This LORA is trained on the Flux Dev model from Black Forest Labs using a Replicate space. For the training data I took a series of photos of flowers focused on the shapes/ forms/ structural elements (slide 1) - with a view to seeing if they would translate into different domains - garments/ architecture etc. 4 iterations, multiple Reddit threads, finally catching up on a workshop delivered by Ismail Seleit, and I feel like I’m making some progress. Outputs (slide 2) range literal interpretations - garments made of flowers, through to garments inspired by the structural elements present in the training images. Love how the shallow depth of field, and rich texture from the images in the training set is also translated. #ai #generativeai #flux

The new Ideogram 2.0 exhibits incredible realism and refined design taste. This new text-to-image AI is an actual competition to Midjourney